Experiencing language in the order that children do: Training on age-ordered child-directed speech facilitates semantic category learning in a recurrent neural network

language acquisition;recurrent neural network, starting small, Elman, CHILDES, child-directed speech, RNN

Huebner, Philip ; Willits ; JonA

关键词: language acquisition;    recurrent neural network, starting small, Elman, CHILDES, child-directed speech, RNN;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/105710/HUEBNER-THESIS-2019.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Previous work has shown that semantic category knowledge can be captured by a distributional learning algorithm operating over naturalistic, noisy child-directed speech (Huebner & Willits, 2018). In chapter 1 of this work, I discuss the algorithm behind this study, and its ability to represent hierarchically organized and abstract knowledge. In chapter 2, I replicate the findings of Huebner & Willits (2018) using a variant of their corpus in which fewer post-processing modifications were applied to the raw transcripts. In chapter 3, I investigate whether training on input in order that children actually experience language provides any learning advantage relative to training in the reverse order Indeed, I found that semantic categorization benefits from training on input which was ordered by the age of the target child compared to input which was ordered in reverse. I refer to this effect as the age-order effect. To investigate what corpus-statistical factors may underlie the age-order effect, I explore structural differences between speech to younger vs. older children in chapter 4. In alignment with previous studies, I found that speech to younger children is syntactically less complex compared to speech to older children. Evidence for differences in semantic category structure was inconsistent. In chapter 5, I propose a number of competing explanations of the age-order effect, and identify one hypothesis, termed the good-start hypothesis, as the most promising. In chapter 6, I expand and refine the good-start hypothesis, and provide further empirical support for it. In chapter 7, I test two core assumptions of the theory developed in chapter 6 using carefully controlled artificial language corpora and find strong support for both. I close with a brief overview of findings in infant behavioral studies consistent with the theory and discuss the implications of the theory for infant acquisition of semantic category knowledge.